2 * Copyright (c) 2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/vfs/hammer/hammer_prune.c,v 1.14 2008/07/11 05:44:23 dillon Exp $
40 * Iterate through the specified range of object ids and remove any
41 * deleted records that fall entirely within a prune modulo.
43 * A reverse iteration is used to prevent overlapping records from being
44 * created during the iteration due to alignments. This also allows us
45 * to adjust alignments without blowing up the B-Tree.
47 static int prune_should_delete(struct hammer_ioc_prune *prune,
48 hammer_btree_leaf_elm_t elm);
49 static void prune_check_nlinks(hammer_cursor_t cursor,
50 hammer_btree_leaf_elm_t elm);
53 hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
54 struct hammer_ioc_prune *prune)
56 struct hammer_cursor cursor;
57 hammer_btree_leaf_elm_t elm;
58 struct hammer_ioc_prune_elm *copy_elms;
59 struct hammer_ioc_prune_elm *user_elms;
65 if (prune->nelms < 0 || prune->nelms > HAMMER_MAX_PRUNE_ELMS)
67 if ((prune->key_beg.localization | prune->key_end.localization) &
68 HAMMER_LOCALIZE_PSEUDOFS_MASK) {
71 if (prune->key_beg.localization > prune->key_end.localization)
73 if (prune->key_beg.localization == prune->key_end.localization) {
74 if (prune->key_beg.obj_id > prune->key_end.obj_id)
76 /* key-space limitations - no check needed */
78 if ((prune->head.flags & HAMMER_IOC_PRUNE_ALL) && prune->nelms)
81 prune->key_cur.localization = prune->key_end.localization +
83 prune->key_cur.obj_id = prune->key_end.obj_id;
84 prune->key_cur.key = HAMMER_MAX_KEY;
87 * Copy element array from userland
89 elm_array_size = sizeof(*copy_elms) * prune->nelms;
90 user_elms = prune->elms;
91 copy_elms = kmalloc(elm_array_size, M_TEMP, M_WAITOK);
92 if ((error = copyin(user_elms, copy_elms, elm_array_size)) != 0)
94 prune->elms = copy_elms;
96 seq = trans->hmp->flusher.act;
99 * Scan backwards. Retries typically occur if a deadlock is detected.
102 error = hammer_init_cursor(trans, &cursor, NULL, NULL);
104 hammer_done_cursor(&cursor);
107 cursor.key_beg.localization = prune->key_beg.localization +
108 ip->obj_localization;
109 cursor.key_beg.obj_id = prune->key_beg.obj_id;
110 cursor.key_beg.key = HAMMER_MIN_KEY;
111 cursor.key_beg.create_tid = 1;
112 cursor.key_beg.delete_tid = 0;
113 cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
114 cursor.key_beg.obj_type = 0;
116 cursor.key_end.localization = prune->key_cur.localization;
117 cursor.key_end.obj_id = prune->key_cur.obj_id;
118 cursor.key_end.key = prune->key_cur.key;
119 cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
120 cursor.key_end.delete_tid = 0;
121 cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
122 cursor.key_end.obj_type = 0;
124 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
125 cursor.flags |= HAMMER_CURSOR_BACKEND;
128 * This flag allows the B-Tree code to clean up loose ends.
130 cursor.flags |= HAMMER_CURSOR_PRUNING;
132 error = hammer_btree_last(&cursor);
138 elm = &cursor.node->ondisk->elms[cursor.index].leaf;
139 prune->key_cur = elm->base;
142 * Yield to more important tasks
144 if ((error = hammer_signal_check(trans->hmp)) != 0)
147 if (prune->stat_oldest_tid > elm->base.create_tid)
148 prune->stat_oldest_tid = elm->base.create_tid;
150 if (hammer_debug_general & 0x0200) {
151 kprintf("check %016llx %016llx cre=%016llx del=%016llx\n",
154 elm->base.create_tid,
155 elm->base.delete_tid);
158 if (prune_should_delete(prune, elm)) {
159 if (hammer_debug_general & 0x0200) {
160 kprintf("check %016llx %016llx: DELETE\n",
161 elm->base.obj_id, elm->base.key);
165 * NOTE: This can return EDEADLK
167 * Acquiring the sync lock guarantees that the
168 * operation will not cross a synchronization
169 * boundary (see the flusher).
171 isdir = (elm->base.rec_type == HAMMER_RECTYPE_DIRENTRY);
173 hammer_sync_lock_sh(trans);
174 error = hammer_delete_at_cursor(&cursor,
175 HAMMER_DELETE_DESTROY,
177 hammer_sync_unlock(trans);
182 ++prune->stat_dirrecords;
184 ++prune->stat_rawrecords;
187 * The current record might now be the one after
188 * the one we deleted, set ATEDISK to force us
189 * to skip it (since we are iterating backwards).
191 cursor.flags |= HAMMER_CURSOR_ATEDISK;
194 * Nothing to delete, but we may have to check other
197 prune_check_nlinks(&cursor, elm);
198 cursor.flags |= HAMMER_CURSOR_ATEDISK;
199 if (hammer_debug_general & 0x0100) {
200 kprintf("check %016llx %016llx: SKIP\n",
201 elm->base.obj_id, elm->base.key);
204 ++prune->stat_scanrecords;
206 if (hammer_flusher_meta_halflimit(trans->hmp) ||
207 hammer_flusher_undo_exhausted(trans, 1)) {
208 hammer_unlock_cursor(&cursor, 0);
209 hammer_flusher_wait(trans->hmp, seq);
210 hammer_lock_cursor(&cursor, 0);
211 seq = hammer_flusher_async(trans->hmp);
213 error = hammer_btree_iterate_reverse(&cursor);
217 hammer_done_cursor(&cursor);
218 if (error == EDEADLK)
220 if (error == EINTR) {
221 prune->head.flags |= HAMMER_IOC_HEAD_INTR;
225 prune->key_cur.localization &= HAMMER_LOCALIZE_MASK;
226 prune->elms = user_elms;
227 kfree(copy_elms, M_TEMP);
232 * Check pruning list. The list must be sorted in descending order.
234 * Return non-zero if the record should be deleted.
237 prune_should_delete(struct hammer_ioc_prune *prune, hammer_btree_leaf_elm_t elm)
239 struct hammer_ioc_prune_elm *scan;
243 * If pruning everything remove all records with a non-zero
246 if (prune->head.flags & HAMMER_IOC_PRUNE_ALL) {
247 if (elm->base.delete_tid != 0)
252 for (i = 0; i < prune->nelms; ++i) {
253 scan = &prune->elms[i];
256 * Check for loop termination.
258 if (elm->base.create_tid >= scan->end_tid ||
259 elm->base.delete_tid > scan->end_tid) {
264 * Determine if we can delete the record.
266 if (elm->base.delete_tid &&
267 elm->base.create_tid >= scan->beg_tid &&
268 elm->base.delete_tid <= scan->end_tid &&
269 (elm->base.create_tid - scan->beg_tid) / scan->mod_tid ==
270 (elm->base.delete_tid - scan->beg_tid) / scan->mod_tid) {
278 * Dangling inodes can occur if processes are holding open descriptors on
279 * deleted files as-of when a machine crashes. When we find one simply
280 * acquire the inode and release it. The inode handling code will then
281 * do the right thing.
285 prune_check_nlinks(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm)
290 if (elm->base.rec_type != HAMMER_RECTYPE_INODE)
292 if (elm->base.delete_tid != 0)
294 if (hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA))
296 if (cursor->data->inode.nlinks)
298 hammer_cursor_downgrade(cursor);
299 ip = hammer_get_inode(cursor->trans, NULL, elm->base.obj_id,
301 elm->base.localization & HAMMER_LOCALIZE_PSEUDOFS_MASK,
304 kprintf("pruning disconnected inode %016llx\n",
306 hammer_rel_inode(ip, 0);
308 kprintf("unable to prune disconnected inode %016llx\n",
316 * NOTE: THIS CODE HAS BEEN REMOVED! Pruning no longer attempts to realign
317 * adjacent records because it seriously interferes with every
318 * mirroring algorithm I could come up with.
320 * This means that historical accesses beyond the first snapshot
321 * softlink should be on snapshot boundaries only. Historical
322 * accesses from "now" to the first snapshot softlink continue to
325 * NOTE: It also looks like there's a bug in the removed code. It is believed
326 * that create_tid can sometimes get set to 0xffffffffffffffff. Just as
327 * well we no longer try to do this fancy shit. Probably the attempt to
328 * correct the rhb is blowing up the cursor's indexing or addressing mapping.
330 * Align the record to cover any gaps created through the deletion of
331 * records within the pruning space. If we were to just delete the records
332 * there would be gaps which in turn would cause a snapshot that is NOT on
333 * a pruning boundary to appear corrupt to the user. Forcing alignment
334 * of the create_tid and delete_tid for retained records 'reconnects'
335 * the previously contiguous space, making it contiguous again after the
338 * The use of a reverse iteration allows us to safely align the records and
339 * related elements without creating temporary overlaps. XXX we should
340 * add ordering dependancies for record buffers to guarantee consistency
344 realign_prune(struct hammer_ioc_prune *prune,
345 hammer_cursor_t cursor, int realign_cre, int realign_del)
347 struct hammer_ioc_prune_elm *scan;
348 hammer_btree_elm_t elm;
353 hammer_cursor_downgrade(cursor);
355 elm = &cursor->node->ondisk->elms[cursor->index];
356 ++prune->stat_realignments;
359 * Align the create_tid. By doing a reverse iteration we guarantee
360 * that all records after our current record have already been
361 * aligned, allowing us to safely correct the right-hand-boundary
362 * (because no record to our right is otherwise exactly matching
363 * will have a create_tid to the left of our aligned create_tid).
366 if (realign_cre >= 0) {
367 scan = &prune->elms[realign_cre];
369 delta = (elm->leaf.base.create_tid - scan->beg_tid) %
372 tid = elm->leaf.base.create_tid - delta + scan->mod_tid;
375 error = hammer_btree_correct_rhb(cursor, tid + 1);
377 error = hammer_btree_extract(cursor,
378 HAMMER_CURSOR_GET_LEAF);
382 error = hammer_cursor_upgrade(cursor);
385 hammer_modify_node(cursor->trans, cursor->node,
386 &elm->leaf.base.create_tid,
387 sizeof(elm->leaf.base.create_tid));
388 elm->leaf.base.create_tid = tid;
389 hammer_modify_node_done(cursor->node);
395 * Align the delete_tid. This only occurs if the record is historical
396 * was deleted at some point. Realigning the delete_tid does not
397 * move the record within the B-Tree but may cause it to temporarily
398 * overlap a record that has not yet been pruned.
400 if (error == 0 && realign_del >= 0) {
401 scan = &prune->elms[realign_del];
403 delta = (elm->leaf.base.delete_tid - scan->beg_tid) %
406 error = hammer_btree_extract(cursor,
407 HAMMER_CURSOR_GET_LEAF);
409 hammer_modify_node(cursor->trans, cursor->node,
410 &elm->leaf.base.delete_tid,
411 sizeof(elm->leaf.base.delete_tid));
412 elm->leaf.base.delete_tid =
413 elm->leaf.base.delete_tid -
414 delta + scan->mod_tid;
415 hammer_modify_node_done(cursor->node);